Problem with sample_node and more_replacements

[twiecki]

X = theano.shared(np.array([1., 2., 3.]))
with pm.Model() as model:
    x = pm.Normal('x', mu=0, sd=1)
    n = pm.Normal('n', mu=X*x, sd=1, observed=np.array([1, 2, 3]))
    
    mean_field = pm.fit(method='advi', n=50000, 
                        callbacks=[pm.callbacks.CheckParametersConvergence(diff='relative')])
    
    out_sample = mean_field.sample_node(n, size=100, more_replacements={X: np.random.randn(1000)})

I expect out_sample.eval().shape to provide shape (100, 1000), but it has shape (100, 3). It seems like the replacement is not correctly applied. This is on pymc3 master.

[junpenglao]

The same problem is discussed here: #2435 i think

[twiecki]

In addition, it seems like sample_node() just returns the exact same value for every sample, thus, uncertainty is not taken into account.

[ferrine]

@twiecki Your example is a bit wrong, you are sampling the observed

[twiecki]

@ferrine Hm, I guess I need to sample the input to the observed?

[ferrine]

And add sigma with white noise

[twiecki]

That seems to work correctly. But it's a bit confusing as sample_ppc samples from the observed.

[DeliciousHair]

I know this is old, but I'm running into the same problem and the conversation above is not giving me any clues.

As an example, if I try to replicate the results in your blog post, I keep hitting this wall here:

with pm.Model() as logistic_regression:
    inputs = pm.Normal('inputs', mu=X, sd=X_unc, shape=X.shape)
    coeffs = pm.Normal('beta', mu=0, sd=1, shape=2)
    
    linear = pm.math.dot(inputs, coeffs)
    prob = pm.math.sigmoid(linear)
    obs = pm.Bernoulli('obs', p=prob, observed=Y)
    
    # Run ADVI which returns posterior means, standard deviations, and the evidence lower bound (ELBO)
    mean_field = pm.fit(method='advi', n=50000,
                        callbacks=[pm.callbacks.CheckParametersConvergence(diff='relative')])
    
    out_sample = (mean_field.sample_node(prob, size=100, more_replacements={inputs: grid_2d}) > .5)

due to the fact that out_sample.eval() has shape (100, 100). Not sure what to make of this...

[ferrine]

Hi! I've replicated your issue in a test suit, going to fix soon